Everyone obsesses over entries. Which indicator. Which model. Which signal threshold. Entries get you into a trade. Exits determine what you keep. I spent more engineering time on my exit system than my entry model. Here’s why — and how it’s structured. Why Exits Are Harder An entry is a single decision: does this bar look like a long or a short? An exit is an ongoing negotiation between multiple objectives: ...
When V5 went live, every single inference run logged the same warning: WARNING: 13 features contain NaN values. Filling with 0. I saw it on day one. I told myself it was a data warmup issue — the live feed just needed more bars to stabilize. That was wrong. Same Code, Different Data V5 solved the training/live code divergence problem from V4. One shared feature_core.py. Every pipeline — training, backtest, all five live scripts — imports from the same file. ...
Day 3. Equity went from $701 to $928. Not because the market was kind. Because the data was finally clean. One line of code. Running for two days. Something felt off on day one. I was holding long positions, the market was moving up, but my equity was bleeding faster than it should. I couldn’t find the cause. On day three, I pulled the logs. Found a keyword repeating every 15 minutes: ...
Every AI session starts with a blank slate. The context window fills up. The session resets. Everything you built together — context, decisions, in-progress work — is gone. Unless you designed around it. The Wrong Way to Handle Session Resets The instinct is to manually copy-paste summaries. Or clear files and reload them. Or re-explain everything at the start of each conversation. That doesn’t scale. Manual handoffs miss things. They compress context incorrectly. And they put the cognitive burden on you — the human — instead of the system. ...
Today I ran a full cross-audit of the V5 quant system. Scope: training, backtest, and live execution — 13 files, 6,500 lines. I was looking for drift. I found something worse. The Trailing Stop Was Updating. It Was Never Checking. My backtest engine has four exit paths: Fixed stop-loss Fixed take-profit Trailing stop Momentum fade SL and TP are delegated to the exchange. The trailing stop has to be managed by the script itself. ...
Most people obsess over model accuracy. Win rate. Precision. Feature importance. That’s 40% of the problem. The other 60% is how you build the backtest that validates it. What the Backtest Actually Controls A model tells you: this bar looks like a long. The backtest decides everything else: How big is the position? Where does the stop go? How does the stop trail? When do you decide the signal is gone? Do you re-enter after an exit? How do you handle overnight funding? Every one of those decisions compounds over hundreds of trades. ...
My V4 trading system used 11 data sources: price data, funding rates, open interest, institutional long/short ratio, liquidation data, fear & greed index, CVD, and more. The backtest results looked solid. Win rates above 80%. Drawdown under 15%. Then I audited the code line by line. What I found made me rebuild the entire system from scratch. Bug 1: The Fear & Greed index was always 50 The Alternative.me API returns data in this format: ...
Most people running a quant system worry about one thing: does the model predict correctly? That’s maybe 30% of the problem. Today I ran a full audit on my BTC and XRP live strategies. Found two serious issues. One system got suspended. Here’s exactly what I checked and what broke. The 4-Dimension Audit Framework After running live for a few days, I realized I needed a structured way to verify the full pipeline — not just “is the model loaded correctly.” ...
Today I pushed three live trading scripts to production. BTC, ETH, SOL. All running simultaneously. 15-minute signals. 10x leverage. Pyramid position sizing up to 5 layers. This is the deployment story. Background I’ve been building this system for months. The short version: Trained LightGBM models on 70 features per symbol Backtested with realistic fees (0.035% with 30% rebate) Ran 70-point audit checks on the live script before touching real money Results: SOL 64% monthly, ETH 56% monthly, BTC 34% monthly (no compounding, $143/symbol) The longer version is in yesterday’s post. ...
I’ll skip the theory. Here’s what actually happened. The Problem I was building a crypto price prediction model for SOL, ETH, and BTC perpetual futures. The plan was simple: train a Transformer on 70 features, get a good signal, trade it live. Three weeks of training later: SOL Transformer v3b: train_acc = 69.3%, val_acc = 54.1%, gap = 15.2% ETH Transformer v3b: train_acc = 78.6%, val_acc = 55.9%, gap = 22.7% The gap between training and validation accuracy was 15-22%. Classic overfitting. The model was memorizing patterns, not learning them. ...